Synthetic data

Synthetic data is data that a computer generates instead of collecting it from real events: realistic but made-up customer records, transactions, photos or texts. Companies use it when real data is scarce, costly or too sensitive to use, such as medical or payment records, for example to test software or to train an AI model. Made-up data is not automatically anonymous, though.

How is it made?

The European Data Protection Supervisor (EDPS) describes synthetic data as artificial data generated from original data and a model trained to reproduce its characteristics and structure, so that both give very similar results in the same statistical analysis. It can also be built from knowledge that analysts supply, or from a mix of both. Common methods:

• Rules and statistics: values are drawn from distributions that someone defines, such as order amounts and delivery dates.
• Generative models: a model such as a GAN learns from real data and produces new examples. MIT Technology Review (February 2022) credits GANs for the boom in synthetic data sets.
• Simulation: Waymo uses simulation to prepare its self-driving software for rare events, train new models and validate new software, and its SimulationCity system synthesises entire journeys (July 2021).
• Language models: a large model writes training examples for another model. With Self-Instruct (Wang et al., 2022), GPT-3 generated instructions and answers that were filtered and used to fine-tune GPT-3 itself. Microsoft trained its code model phi-1 (2023) partly on textbooks and exercises written by GPT-3.5, and NVIDIA reports that over 98% of the data used to align Nemotron-4 340B (2024) was synthetic.

Is synthetic data private?

Not automatically. The EDPS advises a privacy assessment to make sure the synthetic data is not actual personal data, to which the GDPR would apply. Stadler, Oprisanu and Troncoso (USENIX Security 2022) tested a range of generative models and found that synthetic data either did not prevent inference attacks or lost its usefulness, a trade-off no better than traditional anonymisation. Stronger guarantees come from differential privacy, a mathematical limit on what data can reveal about any individual. Google, for example, generates differentially private synthetic text to train safety classifiers (May 2024).

Where does it go wrong?

• It misses outliers. Synthetic data only imitates real data, the EDPS notes, so it may lack the rare cases that matter most in some applications.
• It copies bias. Synthetic data is only as unbiased as the real data used to generate it, MIT Technology Review writes.
• Models can collapse. Shumailov et al. (Nature, July 2024) found that training models on the output of earlier models, generation after generation, makes the rare parts of the original data disappear, which they call model collapse. Gerstgrasser et al. (2024) avoided that collapse by adding synthetic data to the real data instead of replacing it.

Example

Stanford's Alpaca (March 2023) shows both the method and a catch. The researchers had OpenAI's text-davinci-003 generate 52,000 instructions with answers in the style of Self-Instruct, for less than $500, and fine-tuned Meta's LLaMA 7B on them. In their preliminary tests Alpaca behaved much like text-davinci-003. They still ruled out commercial use, partly because OpenAI's terms of use prohibit developing models that compete with OpenAI.